Does dynamic and speculative parallelization enable advanced parallelizing and optimizing code transformations?
Résumé
Thread-Level Speculation (TLS) is a dynamic and automatic parallelization strategy allowing to handle codes that cannot be parallelized at compile-time, because of insufficient information that can be extracted from the source code. However, the proposed TLS systems are strongly limited in the kind of parallelization they can apply on the original sequential code. Consequently, they often yield poor performance. In this paper, we give the main reasons of their limits and show that it is possible in some cases for a TLS system to handle more advanced parallelizing transformations. In particular, it is shown that codes characterized by phases where the memory behavior can be modeled by linear functions, can take advantage of a dynamic use of the polytope model.
Domaines
Informatique et langage [cs.CL]
Origine : Fichiers produits par l'(les) auteur(s)
Loading...